Most AI systems can tell you what they did. Very few can prove it. Cryptographic AI receipts live in that gap.
When a model returns an answer, a routing layer picks a provider, or an agent decides to spend, the record of that event usually lands in a log or a database row the operator controls. That record is a claim. The operator can edit it, lose it, or reconstruct it after the fact, and you have no way to tell. For a demo, that is fine. For anything a buyer, an auditor, or a regulator will later question, a claim is not enough. You need proof. (This is the core distinction between logs and receipts.)
A receipt is a signed record, not a log line
A cryptographic AI receipt captures the conditions of an AI event, things like the model, the inputs, the decision, and the timing, and binds them into a single record signed with a private key. Anyone holding the matching public key can confirm two things at once: the record was produced by the claimed signer, and it has not been changed by so much as a byte since. Change one character and verification fails.
Three properties make it proof rather than record-keeping:
- Signed. Hive receipts are signed with post-quantum ML-DSA-65 (NIST FIPS 204), not stored in a table you have to trust the operator not to touch.
- Verifiable offline. Verification needs only the receipt and the public key. It runs anywhere, any time, with nothing from Hive at check time.
- Third-party checkable. A counterparty confirms the receipt on their own, without your cooperation and without a login.
Why post-quantum, and why offline
Signatures only last as long as the math behind them holds up. A record you might have to defend in three, five, or ten years should not rest on a scheme a future quantum computer could forge after the fact. ML-DSA-65 is the NIST-standardized post-quantum signature (FIPS 204). Sign with it today, and a receipt minted now will still verify, and still mean something, years from now.
Offline verification matters for the same reason. If checking a receipt required calling the platform that made it, the platform would still be the arbiter of its own truth. A receipt you can verify with the public key alone removes that dependency entirely. The proof stands on its own.
What a receipt does not claim
A receipt is honest about its scope. It attests what happened: this model, these inputs, this decision, at this time, not whether the output was correct. It will not tell you an answer was true or a decision was wise. It tells you, verifiably, that the event happened as described and has not been tampered with. That narrow, honest claim is exactly what makes it trustworthy. It never overreaches.
Where receipts start: inference
Inference is the natural place to start, because it is where AI systems make claims most often and prove them least. Hive's AFiR mints a cryptographic receipt for signed AI inference, and for each sub-step along the way, so anyone can verify a model call after the fact. From there, the same primitive extends to agent actions and to human-origin data, all sharing one verifiable record format.
If you want the full picture across all three surfaces (inference, agents, and human data), the Cryptographic AI Receipts overview is the one-page explainer.
Start with the overview, then watch AFiR sign a live inference call and verify it in your browser.